| 赵悦阳,崔雷,王凤婷.判别学习在多模态医学数据融合中的理论应用与创新实践[J].情报工程,2026,(1):051-058 |
| 判别学习在多模态医学数据融合中的理论应用与创新实践 |
| Theoretical Application and Innovative Practice of Discriminative Learning in Multimodal Medical Data Fusion |
| |
| DOI： |
| 中文关键词: 判别学习；多模态医学数据融合；可解释性；精准医疗 |
| 英文关键词: Discriminative Learning; Multimodal Medical Data Fusion; Interpretability; Precision Medicine |
| 基金项目:辽宁省社会科学规划基金资助项目“在多层次相似性共嵌入空间中通过判别学习进行知识发现的描述性文档聚类”( 项目编号：L20BTQ003)。 |
| 作者 | 单位 | | 赵悦阳 | 中国医科大学附属盛京医院图书馆 辽宁沈阳 110004 | | 崔雷 | 中国医科大学健康管理学院 辽宁沈阳 110122 | | 王凤婷 | 中国医科大学附属盛京医院图书馆 辽宁沈阳 110004 |
|
| 摘要点击次数: 4 |
| 全文下载次数: 2 |
| 中文摘要: |
| [目的/意义]本文系统梳理判别学习在多模态医学数据融合中的研究进展，总结其核心方法、技术路径及临床应用潜力，旨在为医学人工智能从感知融合向认知融合的跨越提供理论支撑与实践指导。[方法/过程]通过文献分析与技术归纳，聚焦多模态融合的三大层次——特征级融合、决策级融合以及混合级融合，剖析判别学习在不同融合策略中的创新应用。结合线性判别分析、支持向量机和典型相关分析等传统线性模型，以及自编码器、超图神经网络和Transformer等深度学习方法，探讨特征选择、跨模态对齐与不确定性建模等关键技术，并评估这些方法在癌症生存分析和神经退行性疾病分类等场景中的性能表现。[结果/结论]判别学习通过动态特征权重分配、多模态知识分解及跨尺度协同建模，显著提升了融合模型的准确性与鲁棒性。其在脑疾病诊断、癌症预测等临床场景中的应用已初步验证其价值。[局限]本文介绍的方法在模型复杂性、计算开销与临床可解释性方面仍有待进一步突破。跨模态数据异构性问题尚未完全解决，模型的可解释性仍难以满足医疗实践的严格需求。未来应着力于知识驱动的融合机制、模型轻量化及可信AI 体系的构建。 |
| 英文摘要: |
| [Purpose/Significance] This paper systematically reviews the research progress of discriminative learning in multimodal medical data fusion, summarizes its core methods, technical pathways, and clinical potential, aiming to provide theoretical support and practical guidance for advancing medical artificial intelligence from perceptual fusion to cognitive fusion.[Methods/Process] Through literature analysis and technical synthesis, this work focuses on the three levels of multimodal fusion—feature-level, decision-level, and hybrid-level—and examines the innovative application of discriminative learning across different fusion strategies. It integrates traditional linear models such as linear discriminant analysis, support vector machines, and canonical correlation analysis, along with deep learning approaches including autoencoders, hypergraph neural networks, and Transformers, to explore key techniques including feature selection, cross-modal alignment, and uncertainty
modeling. The performance of these methods is evaluated in scenarios such as cancer survival analysis and neurodegenerative disease classification. [Results/Conclusions] Discriminative learning significantly enhances the accuracy and robustness of fusion models through dynamic feature weight assignment (e.g., attention mechanisms), multimodal knowledge decomposition, and cross-scale collaborative modeling. Its value has been preliminarily validated in clinical scenarios such as brain disease diagnosis and cancer prediction. [Limitations] The methods presented in this paper still require further breakthroughs in model complexity,computational cost, and clinical interpretability. The heterogeneity of cross-modal data remains not fully resolved, and model interpretability still falls short of the stringent requirements of medical practice. Future efforts should focus on knowledge-driven fusion mechanisms, model lightweighting, and the development of trustworthy AI systems. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
